User Max Jendrall demonstrated the ability to deploy a local Gemma 4B language model directly within the ChatGPT isolated environment (sandbox), using unconventional methods to bypass network restrictions.
What Happened
During an experiment, the Gemma 4B model was successfully launched inside the ChatGPT virtual machine. To overcome blocks when downloading weights from Hugging Face, the ChatGPT Sites method was used as a proxy server. Verification confirmed that the computing environment available to paid users possesses resources of 15 GB RAM and 9 CPU cores, which is sufficient for inferencing lightweight models in GGUF format.
Context
Traditionally, LLM sandboxes are perceived as tools for executing code snippets (Code Interpreter) or processing text. However, this technical hack shows that the existing ChatGPT infrastructure can function as a full runtime environment, allowing for the execution of third-party specialized software.
Why It Matters for the Industry
This discovery highlights the potential of cloud LLM sandboxes as full execution environments for autonomous AI agents. Instead of simply generating text, agents could deploy specialized software and manage complex computational processes within a protected cloud perimeter, potentially leading to the concept of AI operating systems.
Why It Matters for Users
For advanced users and developers, this means the ability to use ChatGPT's computing power as a temporary workstation for running small models or performing complex programming tasks, transforming the chatbot from a simple interface into a full-fledged automation tool.
What Is Not Yet Known / Limitations
The method is currently a technical hack and is unsuitable for production use due to the ephemeral nature of the environment, where data and system state are cleared after the session ends, as well as the lack of a stable API to manage this process.
Sources
Author
Look at AI, Editorial Team